deepgram-reference-architecture（Deepgram 参考架构）

Deepgram Reference Architecture

Contents

• Overview
• Prerequisites
• Instructions
• Output
• Error Handling
• Examples
• Resources

Overview

Reference architectures for scalable transcription systems: synchronous API for short files, async queue (BullMQ) for batch processing, WebSocket streaming for real-time, and hybrid routing with enterprise multi-region deployment.

Prerequisites

• Deepgram API access
• Redis for queue-based patterns
• WebSocket support for streaming
• Monitoring infrastructure

Instructions

Step 1: Choose Architecture Pattern

Select based on use case: Sync API for files under 60s, Async Queue for batch/long files, Streaming for real-time transcription, or Hybrid for mixed workloads.

Step 2: Implement Synchronous Pattern

Direct API calls via Express endpoint. Store results in database. Best for low-latency, short audio requirements.

Step 3: Implement Async Queue Pattern

Use BullMQ with Redis for job queuing. Configure workers with concurrency (10), retry (3 attempts, exponential backoff). Notify clients on completion.

Step 4: Implement Streaming Pattern

Create WebSocket server that proxies audio between client and Deepgram Live API. Forward transcripts back to client in real-time with interim results.

Step 5: Build Hybrid Router

Auto-select pattern based on audio duration: sync for <60s, async for >300s. Allow explicit mode override via request parameter.

Step 6: Scale to Enterprise

Deploy multi-region with load balancing. Use Redis cluster for cross-region coordination. Configure per-region worker pools with 20 concurrency and 5 retries.

See detailed implementation for advanced patterns.

Output

• Synchronous transcription endpoint
• Queue-based async processing pipeline
• Real-time WebSocket streaming server
• Hybrid router with auto-selection
• Enterprise multi-region architecture
• Prometheus monitoring integration

Error Handling

Examples

Architecture Selection Guide

Resources

• Deepgram Architecture Guide
• High Availability Patterns
• Scaling Best Practices